Allow disabling USB3 ports in xHCI/DWC3

Hi USB devs,

This patchset is a semi-RFC: I haven't discussed this change yet, and it may turn out to be a bad idea. But if there is a consensus that this change is appropriate, these patches are the ones I'd submit for inclusion.

These patches were developed while working with a SoC (Rockchip RK3588) that contains DWC3-OTG controllers and accompanying USB2 + USB3/DP PHYs. My target (Turing RK1) uses its first bus in USB2.0-OTG mode only: the associated USB3 PHY is reserved for DP. Worse, a driver for the USBDP block (though it exists) has not been merged to mainline. Without lighting up the PHY side of the PIPE, the DWC3 behaves erratically, even for USB2 operation.

This could be addressed by patching in the (out-of-tree) USBDP driver and enabling only its USB backend. However, I found it cleaner (also from a user-friendliness standpoint) just to disable the unusable USB3 port. These patches achieve that by (1) making it possible to tell the xHCI driver to ignore any USB3 port(s), and (2) (perhaps more controversially) making the DWC3 driver disable USB3 host ports when maximum-speed isn't set high enough.

There are other ways to disable the USB3 ports on RK3588, such as via some syscon registers. I figured I would start with the most general solution (benefitting other SoCs) first, getting more specific only if necessary. :)

Thank you for your attention,
Sam

Sam Edwards (2):
xhci: Introduce "disable-usb3" DT property/quirk
usb: dwc3: host: Disable USB3 ports if maximum-speed doesn't permit
USB3

 Documentation/devicetree/bindings/usb/usb-xhci.yaml | 4 ++++
 drivers/usb/dwc3/host.c                             | 5 ++++-
 drivers/usb/host/xhci-mem.c                         | 4 ++++
 drivers/usb/host/xhci-plat.c                        | 3 +++
 drivers/usb/host/xhci.h                             | 1 +
 5 files changed, 16 insertions(+), 1 deletion(-)

2.41.0

Some systems may have xHCI controllers that enumerate USB 3.0 ports, but these ports nevertheless cannot be used. Perhaps enabling them triggers a hardware bug, or perhaps they simply aren't connected and it would be confusing to the user to see an unusable USB 3.0 rhub show up -- whatever the case may be, it's reasonable to want to disable these ports.

Add a DT property (and associated quirk) to the xHCI driver that skips over (i.e. ignores and doesn't initialize) any USB 3.0 ports discovered during driver initialization.

---
 Documentation/devicetree/bindings/usb/usb-xhci.yaml | 4 ++++
 drivers/usb/host/xhci-mem.c                         | 4 ++++
 drivers/usb/host/xhci-plat.c                        | 3 +++
 drivers/usb/host/xhci.h                             | 1 +
 4 files changed, 12 insertions(+)

index 180a261c3e8f..8a64e747260a 100644
@@ -25,6 +25,10 @@ properties:
     description: Set if the controller has broken port disable mechanism
     type: boolean
 
+  disable-usb3:
+    description: Ignore (don't initialize, don't use) USB3 ports
+    type: boolean
+
   imod-interval-ns:
     description: Interrupt moderation interval
     default: 5000

index 0a37f0d511cf..bf8fcab626e4 100644
@@ -1968,6 +1968,10 @@ static void xhci_add_in_port(struct xhci_hcd *xhci, unsigned int num_ports,
 	minor_revision = XHCI_EXT_PORT_MINOR(temp);
 
 	if (major_revision == 0x03) {
+		/* Ignore USB3 ports entirely if USB3 support is disabled. */
+		if (xhci->quirks & XHCI_DISABLE_USB3)
+			return;
+
 		rhub = &xhci->usb3_rhub;
 		/*
 		 * Some hosts incorrectly use sub-minor version for minor

index b93161374293..75285fb5bbbc 100644
@@ -249,6 +249,9 @@ int xhci_plat_probe(struct platform_device *pdev, struct device *sysdev, const s
 		if (device_property_read_bool(tmpdev, "quirk-broken-port-ped"))
 			xhci->quirks |= XHCI_BROKEN_PORT_PED;
 
+		if (device_property_read_bool(tmpdev, "disable-usb3"))
+			xhci->quirks |= XHCI_DISABLE_USB3;
+
 		device_property_read_u32(tmpdev, "imod-interval-ns",
 					 &xhci->imod_interval);
 	}

index 5df370482521..c53fbeea478f 100644
@@ -1906,6 +1906,7 @@ struct xhci_hcd {
 #define XHCI_RESET_TO_DEFAULT	BIT_ULL(44)
 #define XHCI_ZHAOXIN_TRB_FETCH	BIT_ULL(45)
 #define XHCI_ZHAOXIN_HOST	BIT_ULL(46)
+#define XHCI_DISABLE_USB3	BIT_ULL(47)
 
 	unsigned int		num_active_eps;
 	unsigned int		limit_active_eps;

2.41.0

Bindings are always separate patches.

Please do not sneak in properties without DT review.

Please use scripts/get_maintainers.pl to get a list of necessary people and lists to CC. It might happen, that command when run on an older kernel, gives you outdated entries. Therefore please be sure you base your patches on recent Linux kernel.

You missed at least devicetree list (maybe more), so this won't be tested by automated tooling. Performing review on untested code might be a waste of time, thus I will skip this patch entirely till you follow the process allowing the patch to be tested.

Please kindly resend and include all necessary To/Cc entries.

Best regards,
Krzysztof

Hi Krzysztof,

It makes sense that the new property should be introduced in a separate patch. I'll ensure that is the case in v2. (If there is one -- see below.)

I have tried (and failed) to find the documentation for this linux-devicetree bot. Do you have the link? In particular, I'd like to ensure that patch 2/2 (the one that actually changes established behavior) is tested sufficiently thoroughly.

That's fine; this patch has just failed review anyway (due to the new property not being introduced in a separate patch), and I'll need to prepare and send a v2 to proceed. However as I mentioned in the cover, this is a semi-RFC. I haven't discussed the overall idea with anyone yet, so to avoid wasting my own time, I need to give the USB folks ample opportunity to object to the proposed changes or suggest improvements before investing more effort in refining the patchset.

As of now, I'm only seeking commentary, not formal review. I'd appreciate any insights on the approach I've taken and whether there are any potential challenges or alternatives that haven't been explored yet. Therefore, I'll hold off on CC-ing linux-devicetree at this stage to keep the focus on the broader concept, and will loop them in (with any other recipients as appropriate) for v2 when (and if!) there's consensus here on linux-usb that the general direction is worth pursuing.

Happy Saturday,
Sam

It does not really explain why you did not Cc some of the maintainers. If this is a RFC, even though not marked as such in subject prefix, then I guess all maintainers should be involved for comments.

Best regards,
Krzysztof

Hi Krzysztof,

More simply put: I don't want to waste anyone's time by seeking review/commentary from linux-devicetree on a property enabling a feature that someone here on linux-usb might soon deem (summarily) untenable.

I agree fully that I need (and want) the DT reviewers' sign-off on this new property, and I'll get more serious about that (and other things) in v2. I just need to know if there's going to be a v2 first.

Warm regards,
Sam

Am Freitag, 8. Dezember 2023, 22:04:57 CET schrieb Sam Edwards:

I'm very much unsure, where the line goes between hw-quirk and dt-is-not-a-configuration-space - in this specific instance.

DT is meant to describe the actual hardware present and not how any operating system supports it.

So having that usb3phy present in the kernel - even if only in a more limited form as you describe would be my preference.

But for a short-term thing, the usb3-phy in the binding is optional, so so you could "just" deduce the no-usb3 state in your code from its absence from the dt-node?

Heiko

Ahoy Heiko,

Glad to have you in the thread!

I can plumb this quirk through something other than DT if there's an objection to it; however, the intended meaning of this property is indeed hardware-description and not OS-configuration: "This xHC does NOT support USB3. Do not let it tell you otherwise by enumerating USB3 ports; it is mistaken. Attempting to enable the USB3 'ports' will at worst provoke a hardware bug and at best register a usb3 rhub for something that doesn't physically exist."

For v2, I should probably rename the property from something "imperative" ("disable-usb3;") to something "declarative," like "no-usb3;" or "usb3-unconnected;" in order to reflect this intended meaning better. The linux-devicetree reviewers (which I will be looping in with v2) will probably prefer a name like that anyway.

Ah, so while there is a short-term benefit, this is actually meant to be long-term: on the Turing RK1 PCB itself, USB-DWC3 #0 is only wired for USB 2.0, not USB 3.0. The corresponding USBDP PHY pins are connected to (dedicated) DisplayPort pins.

This means there are 4 options:
1. Proceed without enabling that USB-DWC3 at all, giving up USB 2.0.
2. Enable the USBDP in USB3 mode, gaining USB 2.0 but giving up DP.
3. Enable the USBDP in DP mode, having the USB3-PIPE backend initialized (but disconnected from the frontend), as a "pacifier" for the DWC3 so it doesn't start misbehaving.
4. Disable the USB3 ports on the xHC side, by...
a. Telling the xHCI driver that it shouldn't enable USB3 ports. b. Configuring the DWC3 not to enable USB3
(see the RK3588 TRM for usb3otg0_host_u3_port_disable)
c. Configuring the DWC3's number of USB3 ports

      (usb3otg0_host_num_u3_port)

Users won't like #1 or #2, and may be confused by #3 due to the dummy port that shows up. Options #2 and #3 aren't possible until a USBDP driver lands (though you're right that this is only a short-term concern). Option 4a is what I'm currently trying to do because I believe having physically-inaccessible USB3 ports show up in Linux is undesirable in any case, but I can switch to 4b/4c if we want something more SoC-specific.

So (as far as RK3588 is concerned) with option(s) 4(a/b/c), while this does let USB2.0 work before the USBDP driver lands, the "real" benefit is that it requires neither a dummy port nor powering up a block of the chip (the PHY's USB3-specific block) that isn't actually usable/useful because it doesn't connect to a USB 3.0 socket/device.

Though if that still doesn't sound appealing, I'm happy to change up the approach -- I guess the question then is whether it's because you *like* option 3, or *dislike* option 4.

Cheers,
Sam

The DWC3 core can be configured (during IP instantiation, and/or via configuration signals) not to have any USB3 ports. Some SoCs, however, may have USB3 interfaces enabled that do not have USB3 PHYs driving them. This can be due to a few circumstances, including:
a) The hardware designer didn't include USB3 PHYs and neglected to disable the DWC3 core's USB ports (I know of no instance where this has actually happened, however, and it seems pretty unlikely).
b) The USB3 PHYs are present but powered off. Perhaps a driver to enable the PHYs has not yet been written or merged, or USB3 capability is unneeded in the system and the system designer would like to conserve power.
c) The USB3 PHYs are muxed to a different controller. This can happen if the PHYs support non-USB protocols and one of these alternate functions is needed instead.

In these circumstances, if the DWC3 does not receive clear link status indication on an enabled USB3 port, the DWC3 may not allow even USB2 to function: in host mode, the DWC3 generates an endless barrage of PORT_CSC status on the accompanying USB2 port, and the xHCI driver is unable to bring the USB2 port to a functioning state.

Fix this by first checking if the maximum-speed property in the DT permits USB3. If not, pass the new disable-usb3; property to the virtual xHCI device, causing the xHCI driver not to enable the USB3 ports. This allows USB2 to function even with USB3 PHYs missing/misbehaving, and may be useful even when the USB3 PHYs are
well-behaved: a DT author may know that USB3 support is intact, but disconnected (not exposed off-board) and choose to lower the maximum-speed property to avoid an unusable USB3 rhub showing up in sysfs/lsusb where it may mislead end-users.

---
 drivers/usb/dwc3/host.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

index 61f57fe5bb78..29f170927e70 100644
@@ -61,7 +61,7 @@ static int dwc3_host_get_irq(struct dwc3 *dwc)
 
 int dwc3_host_init(struct dwc3 *dwc)
 {
-	struct property_entry	props[4];
+	struct property_entry	props[5];
 	struct platform_device	*xhci;
 	int			ret, irq;
 	int			prop_idx = 0;

@@ -95,6 +95,9 @@ int dwc3_host_init(struct dwc3 *dwc)
 	if (dwc->usb2_lpm_disable)
 		props[prop_idx++] = PROPERTY_ENTRY_BOOL("usb2-lpm-disable");
 
+	if (dwc->maximum_speed == USB_SPEED_FULL || dwc->maximum_speed == USB_SPEED_HIGH)
+		props[prop_idx++] = PROPERTY_ENTRY_BOOL("disable-usb3");
+
 	/**
 	 * WORKAROUND: dwc3 revisions <=3.00a have a limitation
 	 * where Port Disable command doesn't work.

2.41.0

Where is patch 1/2 of this series?

confused,

greg k-h

On 12/15/23 05:44, Greg Kroah-Hartman wrote:> Where is patch 1/2 of this series?

Hi Greg,

I double-checked and it does look like you were Cc'd on it. Either way, it's here:
https://lore.kernel.org/linux-usb/[email protected]/T/#m5203732de487136de9b1d8a93b1c2f0b89dfe5e1

Equally confused,
Sam

I don't think this will work as a generic xhci driver feature.

Even if we ignore all USB3 ports in software they will for most xHC hosts be powered and enabled in hardware by default after controller reset.

This means they perform link training, generate all kinds of events with interrupts (connect, over-current etc) that driver now can't handle.

Sound like the setup you are using has a very specific issue, and it would need a narrow targeted quirk to solve it.

To me a specific solution to a specific problem like this sounds better.

Thanks
Mathias

Hi Mathias,

By this do you mean that having the xHCI driver ignore the USB3 ports isn't enough to ensure that PP=0 (and the driver would need to do a little bit more to make sure that the "parking brake" is on: e.g. initialize, but not use, the ports) or that the xHC's PP=0 signal isn't sufficient to keep the PHYs from trying to bring the link up and generating those interrupts (PP=0 really isn't enough, and there is no general "parking brake" to be found here)?

I infer from this that you're against having a DT property added to xHCI? What if the property were to be narrowed in scope to "ignore the USB3 PHYs, they're disabled/absent" vs. this iteration's "disable the USB3 ports" meaning?

If this quirk ends up landing in the dwc3 driver (since, arguably, DWC3 is the real misbehaving hw block in these circumstances), what would be your preferred mechanism of signaling to the xHCI layer "the USB3 PHYs have been disabled; please ignore"?

I am starting to think so as well. I may shift my focus to DWC3 (with xHCI driver changes made only to facilitate them) for now, since maximum-speed = "high-speed"; very reasonably (imo) should prevent registering the usb3 rhub -- though something may convince me otherwise in the near future. :)

Thanks to you as well, this is exactly the type of feedback I was fishing for!

Cheers,
Sam

Yes, in most cases PP==1 after xHC reset, here's some old debug output during boot:

[    2.571057] xhci_hcd 0000:00:0d.0: new USB bus registered, assigned bus number 2
[    2.571061] xhci_hcd 0000:00:0d.0: Host supports USB 3.2 Enhanced SuperSpeed
[    2.571065] xhci_hcd 0000:00:0d.0: // Turn on HC, cmd = 0x5.
[    2.571067] xhci_hcd 0000:00:0d.0: Finished xhci_run for USB3 roothub
[    2.571093] usb usb2: New USB device found, idVendor=1d6b, idProduct=0003, bcdDevice= 5.15
[    2.571095] usb usb2: New USB device strings: Mfr=3, Product=2, SerialNumber=1
[    2.571097] usb usb2: Product: xHCI Host Controller
[    2.571098] usb usb2: Manufacturer: Linux 5.15.57-06982-gf7339f7585d8-dirty xhci-hcd
[    2.571099] usb usb2: SerialNumber: 0000:00:0d.0
[    2.571197] xHCI xhci_add_endpoint called for root hub
[    2.571199] xHCI xhci_check_bandwidth called for root hub
[    2.571224] hub 2-0:1.0: USB hub found
[    2.571230] hub 2-0:1.0: 2 ports detected
[    2.571279] xhci_hcd 0000:00:0d.0: set port power 2-1 ON, portsc: 0x2a0

Note that portsc: 0x2a0 entry above has PP=1, and it shows the portsc register value _before_ port power is set to 1 (ON).

Port Status: 0x2a0
Disconnected
Disabled
Link: Rx Detect
Powered
Unknown port speed

Forcing PP=0 could possibly prevent any events from those ports.

I don't have a good solution in mind for this so I'll just throw some ideas:

What happends if you in some RK3588 platform code disable USB ports via those syscon registers, but let the xhci driver be, and USB3 roothub enumerates normally?

Or if this is about a misbehaving USB3 PHY, how about adding the USB3 PHY driver that describes reality and fails when usb_phy_roothub_init() or usb_phy_roothub_set_mode() are called by for USB3 hcd during usb_add_hcd(USB3).
xhci driver could then continue without the USB3 hcd. turning off USB3 ports.

Adding the 'maximum-speed = "high-speed"' entry could also be one option.

Thanks
Mathias